Learning with "realistic" synthetic data